Automated soccer head impact exposure tracking using video and deep learning

Area of Science:

• Sports Medicine
• Biomechanical Engineering
• Computer Vision

Background:

• Head impacts in sports are common, necessitating research into their link with brain injury.
• Current methods like wearable sensors and manual video analysis have limitations in cost, accuracy, and time.

Purpose of the Study:

• To develop and validate DeepImpact, a computer vision algorithm for automatic detection of soccer headers from video data.
• To provide a low-cost, efficient alternative for collecting head impact exposure data.

Main Methods:

• A data-driven pipeline using two deep learning networks (object detection and temporal shift module) was developed.
• The algorithm extracts visual and temporal features to classify video segments as headers or non-headers.
• Training and validation were performed on a large-scale professional soccer video dataset.

Main Results:

• DeepImpact achieved 95.3% sensitivity and 96.0% precision in cross-validation.
• In an independent test on five professional games, it reached 92.9% sensitivity and 21.1% precision.
• The algorithm significantly reduces the time required for manual video analysis.

Conclusions:

• DeepImpact offers a streamlined, low-cost, video-based solution for monitoring head impact exposure in soccer.
• This tool can facilitate the collection of large-scale datasets crucial for brain injury research.
• The approach has potential for expansion to other sports requiring head impact analysis.